Upload modeling.py with huggingface_hub

modeling.py

@@ -372,19 +372,18 @@ class NewEmbeddings(nn.Module):
         embeddings = inputs_embeds
 
         # Set and unpad position_ids
-        if position_ids is None:
-            if seq_length > self.position_ids.size(0):
-                self.register_buffer(
-                    "position_ids", torch.arange(seq_length, device=embeddings.device), persistent=False
-                )
-            if unpad_inputs:
-                # [1, cumsum_seq_len]
-                position_ids = torch.cat([self.position_ids[:l] for l in length]).unsqueeze(0)
-            else:
-                # [bs, seq_len]
-                position_ids = self.position_ids[:seq_length].expand(batch_size, -1)
-        elif unpad_inputs:
-            position_ids = position_ids[attention_mask_bool].unsqueeze(0)  # [1, cumsum_seq_len]
+        # Always regenerate internally for robustness — avoids corrupt position_ids
+        # from external callers (e.g. sentence-transformers on some environments)
+        if seq_length > self.position_ids.size(0):
+            self.register_buffer(
+                "position_ids", torch.arange(seq_length, device=embeddings.device), persistent=False
+            )
+        if unpad_inputs:
+            # [1, cumsum_seq_len]
+            position_ids = torch.cat([self.position_ids[:l] for l in length]).unsqueeze(0)
+        else:
+            # [bs, seq_len]
+            position_ids = self.position_ids[:seq_length].expand(batch_size, -1)
 
         # Compute rotary embedding
         if self.position_embedding_type == 'rope':